ChatBCI: A P300 Speller BCI Leveraging Large Language Models for Improved Sentence Composition in Realistic Scenarios

P300 speller BCIs allow users to compose sentences by selecting target keys on a GUI through the detection of P300 component in their EEG signals following visual stimuli. Most P300 speller BCIs require users to spell words letter by letter, or the first few initial letters, resulting in high keystroke demands that increase time, cognitive load, and fatigue. This highlights the need for more efficient, user-friendly methods for faster sentence composition. In this work, we introduce ChatBCI, a P300 speller BCI that leverages the zero-shot learning capabilities of large language models (LLMs) to suggest words from user-spelled initial letters or predict the subsequent word(s), reducing keystrokes and accelerating sentence composition. ChatBCI retrieves word suggestions through remote queries to the GPT-3.5 API. A new GUI, displaying GPT-3.5 word suggestions as extra keys is designed. SWLDA is used for the P300 classification. Seven subjects completed two online spelling tasks: 1) copy-spelling a self-composed sentence using ChatBCI, and 2) improvising a sentence using ChatBCI’s word suggestions. Results demonstrate that in Task 1, on average, ChatBCI outperforms letter-by-letter BCI spellers, reducing time and keystrokes by 62.14% and 53.22%, respectively, and increasing information transfer rate by 198.96%. In Task 2, ChatBCI achieves 80.68% keystroke savings and a record 8.53 characters/min for typing speed. Overall, ChatBCI, by employing remote LLM queries, enhances sentence composition in realistic scenarios, significantly outperforming traditional spellers without requiring local model training or storage. ChatBCI’s (multi-) word predictions, combined with its new GUI, pave the way for developing next-generation speller BCIs that are efficient and effective for real-time communication, especially for users with communication and motor disabilities.

💡 Research Summary

The paper introduces ChatBCI, a novel P300‑based brain‑computer interface speller that integrates the GPT‑3.5‑turbo large language model (LLM) via remote API calls to provide word‑completion and next‑word prediction capabilities. Traditional P300 spellers rely on a 6 × 6 (or similar) matrix where rows and columns flash repeatedly; users select a target key by focusing on it, eliciting a P300 ERP approximately 300 ms after the deviant stimulus. While many studies have improved flashing patterns, stimulus intervals, or classification algorithms, they still require users to spell each character, leading to high keystroke counts, long task times, and increased cognitive load—especially for composing full sentences.

ChatBCI addresses this limitation by sending the partially typed text to the GPT‑3.5 API using a carefully engineered prompt template. The prompt consists of two messages: a system‑role instruction that tells the model to act as a “professional spelling and grammar corrector” and to return a maximum of ten candidate words, and a user‑role message containing the current partial sentence. The model’s response is a plain list of candidate words, which the system parses and displays as ten extra keys on the left and right sides of a 5 × 8 keyboard matrix. The GUI therefore shows 26 alphabetic keys, four functional keys (delete word, delete character, space, and enter), and ten suggestion slots that update in real time as the user types.

Signal acquisition follows the classic row‑column flashing paradigm. EEG is recorded, filtered, and processed online; Stepwise Linear Discriminant Analysis (SW‑LDA) classifies each flash as target or non‑target, determining the selected key. Once a key is selected, the system updates the displayed sentence and immediately issues a new GPT‑3.5 query based on the updated text, ensuring continuous, context‑aware suggestions.

The authors evaluated ChatBCI with seven healthy participants across two online tasks. Task 1 required participants to copy a pre‑defined sentence (copy‑spelling), while Task 2 asked them to improvise a sentence using the suggestions (improvisation). Results for Task 1 showed an average keystroke reduction of 62.14 % and a time reduction of 53.22 % compared with a conventional letter‑by‑letter P300 speller, yielding an information transfer rate (ITR) increase of 198.96 %. In the more naturalistic improvisation task, participants saved 80.68 % of keystrokes and achieved a record typing speed of 8.53 characters per minute. These figures demonstrate that remote LLM integration can dramatically accelerate sentence composition without the need for local model training, storage, or inference hardware.

The discussion acknowledges several practical considerations. Remote API calls introduce latency, which could affect real‑time responsiveness; the authors suggest caching or fallback lightweight models for offline use. Privacy concerns arise when user‑generated text is transmitted to an external service, prompting the need for encryption or on‑device anonymization. Moreover, GPT‑3.5 may occasionally propose words that are syntactically plausible but semantically inappropriate for the user’s intent; a feedback loop or confidence‑based filtering could mitigate this. The current system limits suggestions to ten candidates; expanding this number or adapting the prompt dynamically based on user behavior could further improve efficiency.

In conclusion, ChatBCI demonstrates that leveraging zero‑shot capabilities of large language models via simple API queries can transform P300 spellers from character‑level input devices into context‑aware, predictive communication tools. By eliminating the overhead of training and maintaining local language models, the approach offers a scalable pathway toward practical, high‑performance assistive communication technologies for individuals with motor or speech impairments. Future work will explore multi‑language support, adaptive prompting, and tighter integration of language model confidence scores into the BCI decision pipeline.